Back to search

University of Ioannina

Knowledge extraction methods from document collections

Abstract

dc:description

This thesis studies the problem of document clustering. Given a document collection, at first, preprocessing, and feature extraction take place. As a result, each document is usually represented using a vector space model where the non-negative dimension weights describe the significance of the respective term features. The properties of such a feature space are: i) the high dimensionality that is of the order of thousands of features, and ii) sparsity which reaches 99%. In this dissertation, methods are studied and developed for document representation and knowledge extraction regarding the cluster structure of a dataset. At first, a vector space model is presented which, without supervision, revisits the traditional assumption about the term independence. A Global Term Context Vector is computed for each term feature of the collection, which embeds the context in which a term appears in the documents (term co-occurrences). Next, a semantic matrix is constructed based on which the document vectors are mapped in a denser feature space of the same dimensionality. The effectiveness of the proposed representation model is experimentally studied in the context of document clustering. The second contribution of this work is the k-synthetic prototypes clustering method that is based on the spherical k-means. Its novelty lays at the introduction of the synthetic prototypes as cluster representatives. The proposed incremental approach for the computation of a synthetic prototype uses the K nearest neighbors of the cluster medoid. The interesting property of this approach is that it favors the representation of the documents of the dominant class in the cluster. In this way the clustering algorithm manages to overcome problems caused by bad initializations. In the experimental study, this method is compared to a series of widely-used document clustering techniques. In the chapter that follows, incremental clustering algorithms are studied, which add the k+1 data based on a solution containing k clusters. A general framework for incremental clustering is presented which applies partial update of the solution when introducing the k+1 cluster. This framework covers known incremental algorithms, such as bisecting k-means, global k-means, and various extensions of them. Next, global k-synthetic prototypes algorithm is proposed which is experimentally compared to existing incremental approaches achieving better clustering results on document collections. The next chapter concerns the problem of estimating the number of clusters in a dataset. Dip-dist criterion is proposed which considers each object of a cluster as a 'viewer ' and applies a univariate statistic hypothesis test, the dip-test, to examine unimodality in the distribution of the distances between the viewer and the rest of the objects in cluster. This criterion is incorporated by the incremental dip-means method. The only assumption of this method is the unimodality of all clusters. Important advantages are: i) the unimodality test is applied on univariate distance vectors, and ii) it can be directly applied with kernel-based methods, since only the pairwise distances are involved in the computations. Experimental results on artificial and real datasets indicate the effectiveness of our method and its superiority over analogous approaches.

Degree

thesis:*
Grantor dc:publisher
University of Ioannina
Year dc:date
2013

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kalogeratos, Argyris
  • Καλογεράτος, Αργύρης

Subjects

dc:subject × 16

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/29374

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Kalogeratos, Argyris; Καλογεράτος, Αργύρης. Knowledge extraction methods from document collections. University of Ioannina, 2013. http://hdl.handle.net/10442/hedi/29374